Re: git status always modifies index?

From: Jeff King <peff@peff.net>
Date: Nov 22, 2017, 21:17 UTC
Message-ID: <20171122211729.GA2854@sigill>
In-Reply-To: <20171122202720.GD11671@aiede.mtv.corp.google.com>

On Wed, Nov 22, 2017 at 12:27:20PM -0800, Jonathan Nieder wrote:

Show 19 quoted lines

> Nathan Neulinger wrote[1]:
> 
> > I just got an answer to my stackoverflow question on this,
> > apparently it's already implemented:
> >
> > https://stackoverflow.com/questions/47436939/how-to-run-git-status-without-modifying-git-index-such-as-in-a-prompt-command
> >
> > There is a "--no-optional-locks" command in 2.15 that looks like it
> > does exactly what I need.
> 
> I was about to point to
> https://public-inbox.org/git/20170921043214.pyhdsrpy4omy54rm@sigill.intra.peff.net/
> about exactly this thing. :)
> 
> That said, I wonder if this use case is an illustration that a name
> like --no-lock-index (as was used in Git for Windows when this feature
> first appeared) or --no-refresh-on-disk-index (sorry, I am terrible at
> coming up with option names) would make the feature easier to
> discover.

Yeah, it's interesting that Nathan does not care about the simultaneous locking here, but rather about the effect of writing to the repo for what would otherwise be a read-only operation.

Under the original intent of --no-optional-locks I think if we could somehow magically update the on-disk index without lock contention, it would be OK to do so. But that would make it no longer work for this particular case.

And I would also not be surprised if there are other cases where we write in a lockless way that would best be avoided in a multi-user setup. I'm thinking specifically of the way that some merge-y operations may write out intermediate objects, even though they're only needed inside the process. It _should_ be a read-only operation to ask "can these two things be merged cleanly", and you should be able to ask that without accidentally creating root-owned files in .git/objects.

So I actually think what Nathan wants is not exactly the same as --no-optional-locks in the first place. But in practice, for a limited set of operations and with the way that locks work in Git, it accomplishes the same thing. Maybe that points to having a broader option. Or maybe having two separate options that largely have the same effect. Or maybe just living with the minor philosophical rough edges, since it seems OK in practice.

-Peff